Carbix Corp + Xyris XR: From Virtual Installation to Industrial Digital Twin

Bringing Carbon Capture, AI, Adaptive Learning and Digital Twins into the Virtual Industrial Environment

One of the most powerful industrial XR projects I’ve worked on was my collaboration with Carbix Corp at their invitation to bring Carbix’s proprietary X2 carbon-capture technology into a fully interactive virtual environment.

Carbix MVSE Highlight reel: https://youtu.be/gXXlpE1EMhM?si=W7D79yr2If9W-9SD

Carbix Corp’s mission is visionary & practical: capture industrial and atmospheric CO₂ and transform it into useful products, including building materials, additives, and valuable gases. The X2 is therefore not an abstract technology - its working prototype is already far into testing. The X2 is flexible on where and how it is deployed—inside the complex physical environments of industries such as cement, float glass and geothermal energy at emission points or as an atmospheric processor.

That made the project an ideal opportunity to explore something I believe is increasingly important in 2026 as wildfires decimate landscapes and property: how do we make complex industrial technology understandable, explorable and trainable before—or alongside—its physical deployment?

Rather than simply creating a 3D model of the X2, we created a virtual industrial installation guided by conversational avatars.

From Product Model to Virtual Installation

We digitally installed the Carbix X2 into three distinct industrial environments: a geothermal caldera, a float-glass factory and a cement plant.

Each environment presented a different context for understanding the technology. The X2 could be examined as part of an operating industrial system rather than as an isolated piece of equipment.

We also created a dedicated X2 showroom, where the technology could be presented through product animations, guided explanations and interactive demonstrations.

This distinction is important. A conventional product visualization answers the question, “What does this machine look like?”

A virtual installation begins to answer much more useful questions:

How is it transported? How does it interact with the facility? Who operates it? What does the surrounding process look like? What does a worker need to know? What happens when something changes?

That is where we began to build the virtual environment not simply as marketing, but as a new kind of industrial interface.

Building the Human Layer

The next step was making the environment something people could actually learn from.

We created a team of conversational avatars representing different perspectives within the Carbix operation. Anya-01 became the primary host, while sales, safety, site-management and operational characters contributed their own viewpoints on carbon capture and the X2 as visitors explore the environment.

I wanted these characters to feel more like participants in an environment than animated information panels. That meant paying particular attention to facial expression, gestures, body language and other non-verbal communication.

We also introduced CORA, a small floating robotic guide that accompanies the visitor through the experience. CORA guides the visitor to points of interest throughout each installation location.

Behind the characters, we integrated speech recognition, text-to-speech and LLM-based conversation, creating an early example of what I now think of as the AI-native industrial environment: a virtual space where information is collated, calibrated, displayed, and discussed.

That distinction becomes more relevant as conversational AI and agentic systems evolve. The industrial interface of the future does not necessarily need to be only another dashboard. It can be an environment populated by intelligent interfaces that understand the equipment, the process and the person using them.

Adaptive Learning

The Carbix MVSE also gave us the opportunity to go beyond conversational AI and investigate adaptive learning.

With support from SingularityNET DeepFunding, we developed a proof-of-concept adaptive learning system based on the SciQA dataset. Rather than giving every learner exactly the same sequence of questions, the system evaluated performance and dynamically selected subsequent questions according to the learner’s demonstrated understanding.

A human instructor naturally adjusts a lesson. If a student understands a concept, the instructor can increase its complexity. If they are struggling, the instructor can slow down, repeat an important concept or approach it from another direction.

We wanted the virtual instructor to be capable of similar guidance.

The research explored a recommendation system capable of navigating relationships between lesson knowledge and learner performance, with the longer-term possibility of incorporating spoken responses, text and eventually signed communication.

This was an important transition for me: from building virtual environments that contain information to building environments that can respond to what the learner actually understands.

Accessibility as Infrastructure

Accessibility was another area where the project pushed the technology forward.

We began development of an ASL recognition system with the intention of making industrial virtual environments accessible through American Sign Language. This was not treated as an optional interface feature. If XR is going to become a meaningful medium for industrial education and training, then communication needs to work for people with different abilities to hear, speak and interact. In turn, software interfaces become more usable by sighted, hearing, mobile students as well.

ASL also reinforced something I had already learned from working on avatar communication: language is not just words. Expression, posture, gesture, attention and timing all contribute to meaning.

That made the work on naturalistic avatar behaviour particularly valuable to us. The same non-verbal systems that make an AI character feel more believable can also become part of a more accessible communication system.

TwinCore™: From Virtual Installation to Digital Twin Thinking

The Carbix MVSE project ultimately led me toward a larger question: if we can put a real industrial technology into a virtual facility, populate that facility with people and AI, connect learning to it, and eventually connect it to operational information—where does the virtual installation end and the digital twin begin?

That question led to my subsequent independent project, TwinCore™.

TwinCore™ became a testbed for exploring the industrial environment at a much larger systems level. Instead of focusing on one machine, I began modelling relationships between equipment, infrastructure, operations, logistics, maintenance, environmental conditions, feedstock supply issues and business factors.

The emphasis shifted from “How do we represent this object?” to “How do we represent the relationships between all these objects, influenced by rapidly changing inputs?”

That is the essence of a system-of-systems approach.

TwinCore™ allows me to experiment with asset relationships, equipment health, operational scenarios, logistics, inspection concepts and cascading effects. Today it works with synthetic data, but the architecture is designed around the possibility of integrating live operational data, IoT systems and intelligent analysis.

It is deliberately not a replacement for engineering or production-control systems. Instead, it is a visualization and simulation layer—a place where complex industrial relationships can be made understandable and where new ideas around predictive maintenance, visual inspection and AI-assisted decision support can be explored.

Why This Matters in 2026

In 2026, we are seeing AI move rapidly from a tool that generates content toward agentic systems that can interpret information, converse with users, reason across data and increasingly act as interfaces to other systems. At the same time, digital twins are moving beyond static 3D representations toward models of operational relationships and behaviour linked to real-world infrastructure inventory.

Xyris XR provides a particularly compelling interface between those technologies and people.

A technician can walk through a facility. An AI avatar can accompany data display with the status a piece of equipment. A learner can be assessed and guided according to their level of understanding. A digital twin can expose relationships that would otherwise remain buried in complexity. Operational data can eventually update the environment. Simulation can test scenarios before they are encountered in the physical world.

That is the direction I see industrial XR moving.

Our Carbix MVSE collaboration demonstrated the first part of that journey: taking a real industrial technology and placing it into meaningful virtual contexts. Adaptive learning and conversational avatars added intelligence and human interaction. ASL research began opening the experience to more users. TwinCore™ then extended the idea from a virtual installation toward a broader model of interconnected industrial systems.

For me, the most important accomplishment wasn't any individual model, avatar or AI feature. It was discovering a methodology for bringing them together.

Build the environment. Understand the system. Reveal compexity. Give people intelligent ways to interact with it. Then connect the environment to data, simulation and AI.

That is how a virtual industrial installation can evolve from a presentation into a platform for learning, experimentation, operational understanding and, ultimately, the digital twin.